1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
16 //===----------------------------------------------------------------------===//
18 #include "llvm/Transforms/Scalar.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/Hashing.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/PostOrderIterator.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AssumptionCache.h"
29 #include "llvm/Analysis/CFG.h"
30 #include "llvm/Analysis/ConstantFolding.h"
31 #include "llvm/Analysis/InstructionSimplify.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/MemoryBuiltins.h"
34 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
35 #include "llvm/Analysis/PHITransAddr.h"
36 #include "llvm/Analysis/ValueTracking.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/Dominators.h"
39 #include "llvm/IR/GlobalVariable.h"
40 #include "llvm/IR/IRBuilder.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/LLVMContext.h"
43 #include "llvm/IR/Metadata.h"
44 #include "llvm/IR/PatternMatch.h"
45 #include "llvm/Support/Allocator.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Analysis/TargetLibraryInfo.h"
49 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 #include "llvm/Transforms/Utils/Local.h"
51 #include "llvm/Transforms/Utils/SSAUpdater.h"
54 using namespace PatternMatch;
56 #define DEBUG_TYPE "gvn"
58 STATISTIC(NumGVNInstr, "Number of instructions deleted");
59 STATISTIC(NumGVNLoad, "Number of loads deleted");
60 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
61 STATISTIC(NumGVNBlocks, "Number of blocks merged");
62 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
63 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
64 STATISTIC(NumPRELoad, "Number of loads PRE'd");
66 static cl::opt<bool> EnablePRE("enable-pre",
67 cl::init(true), cl::Hidden);
68 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
70 // Maximum allowed recursion depth.
71 static cl::opt<uint32_t>
72 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
73 cl::desc("Max recurse depth (default = 1000)"));
75 //===----------------------------------------------------------------------===//
77 //===----------------------------------------------------------------------===//
79 /// This class holds the mapping between values and value numbers. It is used
80 /// as an efficient mechanism to determine the expression-wise equivalence of
86 SmallVector<uint32_t, 4> varargs;
88 Expression(uint32_t o = ~2U) : opcode(o) { }
90 bool operator==(const Expression &other) const {
91 if (opcode != other.opcode)
93 if (opcode == ~0U || opcode == ~1U)
95 if (type != other.type)
97 if (varargs != other.varargs)
102 friend hash_code hash_value(const Expression &Value) {
103 return hash_combine(Value.opcode, Value.type,
104 hash_combine_range(Value.varargs.begin(),
105 Value.varargs.end()));
110 DenseMap<Value*, uint32_t> valueNumbering;
111 DenseMap<Expression, uint32_t> expressionNumbering;
113 MemoryDependenceAnalysis *MD;
116 uint32_t nextValueNumber;
118 Expression create_expression(Instruction* I);
119 Expression create_cmp_expression(unsigned Opcode,
120 CmpInst::Predicate Predicate,
121 Value *LHS, Value *RHS);
122 Expression create_extractvalue_expression(ExtractValueInst* EI);
123 uint32_t lookup_or_add_call(CallInst* C);
125 ValueTable() : nextValueNumber(1) { }
126 uint32_t lookup_or_add(Value *V);
127 uint32_t lookup(Value *V) const;
128 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
129 Value *LHS, Value *RHS);
130 void add(Value *V, uint32_t num);
132 void erase(Value *v);
133 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
134 AliasAnalysis *getAliasAnalysis() const { return AA; }
135 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
136 void setDomTree(DominatorTree* D) { DT = D; }
137 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
138 void verifyRemoved(const Value *) const;
143 template <> struct DenseMapInfo<Expression> {
144 static inline Expression getEmptyKey() {
148 static inline Expression getTombstoneKey() {
152 static unsigned getHashValue(const Expression e) {
153 using llvm::hash_value;
154 return static_cast<unsigned>(hash_value(e));
156 static bool isEqual(const Expression &LHS, const Expression &RHS) {
163 //===----------------------------------------------------------------------===//
164 // ValueTable Internal Functions
165 //===----------------------------------------------------------------------===//
167 Expression ValueTable::create_expression(Instruction *I) {
169 e.type = I->getType();
170 e.opcode = I->getOpcode();
171 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
173 e.varargs.push_back(lookup_or_add(*OI));
174 if (I->isCommutative()) {
175 // Ensure that commutative instructions that only differ by a permutation
176 // of their operands get the same value number by sorting the operand value
177 // numbers. Since all commutative instructions have two operands it is more
178 // efficient to sort by hand rather than using, say, std::sort.
179 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
180 if (e.varargs[0] > e.varargs[1])
181 std::swap(e.varargs[0], e.varargs[1]);
184 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
185 // Sort the operand value numbers so x<y and y>x get the same value number.
186 CmpInst::Predicate Predicate = C->getPredicate();
187 if (e.varargs[0] > e.varargs[1]) {
188 std::swap(e.varargs[0], e.varargs[1]);
189 Predicate = CmpInst::getSwappedPredicate(Predicate);
191 e.opcode = (C->getOpcode() << 8) | Predicate;
192 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
193 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
195 e.varargs.push_back(*II);
201 Expression ValueTable::create_cmp_expression(unsigned Opcode,
202 CmpInst::Predicate Predicate,
203 Value *LHS, Value *RHS) {
204 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
205 "Not a comparison!");
207 e.type = CmpInst::makeCmpResultType(LHS->getType());
208 e.varargs.push_back(lookup_or_add(LHS));
209 e.varargs.push_back(lookup_or_add(RHS));
211 // Sort the operand value numbers so x<y and y>x get the same value number.
212 if (e.varargs[0] > e.varargs[1]) {
213 std::swap(e.varargs[0], e.varargs[1]);
214 Predicate = CmpInst::getSwappedPredicate(Predicate);
216 e.opcode = (Opcode << 8) | Predicate;
220 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
221 assert(EI && "Not an ExtractValueInst?");
223 e.type = EI->getType();
226 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
227 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
228 // EI might be an extract from one of our recognised intrinsics. If it
229 // is we'll synthesize a semantically equivalent expression instead on
230 // an extract value expression.
231 switch (I->getIntrinsicID()) {
232 case Intrinsic::sadd_with_overflow:
233 case Intrinsic::uadd_with_overflow:
234 e.opcode = Instruction::Add;
236 case Intrinsic::ssub_with_overflow:
237 case Intrinsic::usub_with_overflow:
238 e.opcode = Instruction::Sub;
240 case Intrinsic::smul_with_overflow:
241 case Intrinsic::umul_with_overflow:
242 e.opcode = Instruction::Mul;
249 // Intrinsic recognized. Grab its args to finish building the expression.
250 assert(I->getNumArgOperands() == 2 &&
251 "Expect two args for recognised intrinsics.");
252 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
253 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
258 // Not a recognised intrinsic. Fall back to producing an extract value
260 e.opcode = EI->getOpcode();
261 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
263 e.varargs.push_back(lookup_or_add(*OI));
265 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
267 e.varargs.push_back(*II);
272 //===----------------------------------------------------------------------===//
273 // ValueTable External Functions
274 //===----------------------------------------------------------------------===//
276 /// add - Insert a value into the table with a specified value number.
277 void ValueTable::add(Value *V, uint32_t num) {
278 valueNumbering.insert(std::make_pair(V, num));
281 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
282 if (AA->doesNotAccessMemory(C)) {
283 Expression exp = create_expression(C);
284 uint32_t &e = expressionNumbering[exp];
285 if (!e) e = nextValueNumber++;
286 valueNumbering[C] = e;
288 } else if (AA->onlyReadsMemory(C)) {
289 Expression exp = create_expression(C);
290 uint32_t &e = expressionNumbering[exp];
292 e = nextValueNumber++;
293 valueNumbering[C] = e;
297 e = nextValueNumber++;
298 valueNumbering[C] = e;
302 MemDepResult local_dep = MD->getDependency(C);
304 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
305 valueNumbering[C] = nextValueNumber;
306 return nextValueNumber++;
309 if (local_dep.isDef()) {
310 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
312 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
313 valueNumbering[C] = nextValueNumber;
314 return nextValueNumber++;
317 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
318 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
319 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
321 valueNumbering[C] = nextValueNumber;
322 return nextValueNumber++;
326 uint32_t v = lookup_or_add(local_cdep);
327 valueNumbering[C] = v;
332 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
333 MD->getNonLocalCallDependency(CallSite(C));
334 // FIXME: Move the checking logic to MemDep!
335 CallInst* cdep = nullptr;
337 // Check to see if we have a single dominating call instruction that is
339 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
340 const NonLocalDepEntry *I = &deps[i];
341 if (I->getResult().isNonLocal())
344 // We don't handle non-definitions. If we already have a call, reject
345 // instruction dependencies.
346 if (!I->getResult().isDef() || cdep != nullptr) {
351 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
352 // FIXME: All duplicated with non-local case.
353 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
354 cdep = NonLocalDepCall;
363 valueNumbering[C] = nextValueNumber;
364 return nextValueNumber++;
367 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
368 valueNumbering[C] = nextValueNumber;
369 return nextValueNumber++;
371 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
372 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
373 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
375 valueNumbering[C] = nextValueNumber;
376 return nextValueNumber++;
380 uint32_t v = lookup_or_add(cdep);
381 valueNumbering[C] = v;
385 valueNumbering[C] = nextValueNumber;
386 return nextValueNumber++;
390 /// lookup_or_add - Returns the value number for the specified value, assigning
391 /// it a new number if it did not have one before.
392 uint32_t ValueTable::lookup_or_add(Value *V) {
393 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
394 if (VI != valueNumbering.end())
397 if (!isa<Instruction>(V)) {
398 valueNumbering[V] = nextValueNumber;
399 return nextValueNumber++;
402 Instruction* I = cast<Instruction>(V);
404 switch (I->getOpcode()) {
405 case Instruction::Call:
406 return lookup_or_add_call(cast<CallInst>(I));
407 case Instruction::Add:
408 case Instruction::FAdd:
409 case Instruction::Sub:
410 case Instruction::FSub:
411 case Instruction::Mul:
412 case Instruction::FMul:
413 case Instruction::UDiv:
414 case Instruction::SDiv:
415 case Instruction::FDiv:
416 case Instruction::URem:
417 case Instruction::SRem:
418 case Instruction::FRem:
419 case Instruction::Shl:
420 case Instruction::LShr:
421 case Instruction::AShr:
422 case Instruction::And:
423 case Instruction::Or:
424 case Instruction::Xor:
425 case Instruction::ICmp:
426 case Instruction::FCmp:
427 case Instruction::Trunc:
428 case Instruction::ZExt:
429 case Instruction::SExt:
430 case Instruction::FPToUI:
431 case Instruction::FPToSI:
432 case Instruction::UIToFP:
433 case Instruction::SIToFP:
434 case Instruction::FPTrunc:
435 case Instruction::FPExt:
436 case Instruction::PtrToInt:
437 case Instruction::IntToPtr:
438 case Instruction::BitCast:
439 case Instruction::Select:
440 case Instruction::ExtractElement:
441 case Instruction::InsertElement:
442 case Instruction::ShuffleVector:
443 case Instruction::InsertValue:
444 case Instruction::GetElementPtr:
445 exp = create_expression(I);
447 case Instruction::ExtractValue:
448 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
451 valueNumbering[V] = nextValueNumber;
452 return nextValueNumber++;
455 uint32_t& e = expressionNumbering[exp];
456 if (!e) e = nextValueNumber++;
457 valueNumbering[V] = e;
461 /// lookup - Returns the value number of the specified value. Fails if
462 /// the value has not yet been numbered.
463 uint32_t ValueTable::lookup(Value *V) const {
464 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
465 assert(VI != valueNumbering.end() && "Value not numbered?");
469 /// lookup_or_add_cmp - Returns the value number of the given comparison,
470 /// assigning it a new number if it did not have one before. Useful when
471 /// we deduced the result of a comparison, but don't immediately have an
472 /// instruction realizing that comparison to hand.
473 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
474 CmpInst::Predicate Predicate,
475 Value *LHS, Value *RHS) {
476 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
477 uint32_t& e = expressionNumbering[exp];
478 if (!e) e = nextValueNumber++;
482 /// clear - Remove all entries from the ValueTable.
483 void ValueTable::clear() {
484 valueNumbering.clear();
485 expressionNumbering.clear();
489 /// erase - Remove a value from the value numbering.
490 void ValueTable::erase(Value *V) {
491 valueNumbering.erase(V);
494 /// verifyRemoved - Verify that the value is removed from all internal data
496 void ValueTable::verifyRemoved(const Value *V) const {
497 for (DenseMap<Value*, uint32_t>::const_iterator
498 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
499 assert(I->first != V && "Inst still occurs in value numbering map!");
503 //===----------------------------------------------------------------------===//
505 //===----------------------------------------------------------------------===//
509 struct AvailableValueInBlock {
510 /// BB - The basic block in question.
513 SimpleVal, // A simple offsetted value that is accessed.
514 LoadVal, // A value produced by a load.
515 MemIntrin, // A memory intrinsic which is loaded from.
516 UndefVal // A UndefValue representing a value from dead block (which
517 // is not yet physically removed from the CFG).
520 /// V - The value that is live out of the block.
521 PointerIntPair<Value *, 2, ValType> Val;
523 /// Offset - The byte offset in Val that is interesting for the load query.
526 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
527 unsigned Offset = 0) {
528 AvailableValueInBlock Res;
530 Res.Val.setPointer(V);
531 Res.Val.setInt(SimpleVal);
536 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
537 unsigned Offset = 0) {
538 AvailableValueInBlock Res;
540 Res.Val.setPointer(MI);
541 Res.Val.setInt(MemIntrin);
546 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
547 unsigned Offset = 0) {
548 AvailableValueInBlock Res;
550 Res.Val.setPointer(LI);
551 Res.Val.setInt(LoadVal);
556 static AvailableValueInBlock getUndef(BasicBlock *BB) {
557 AvailableValueInBlock Res;
559 Res.Val.setPointer(nullptr);
560 Res.Val.setInt(UndefVal);
565 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
566 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
567 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
568 bool isUndefValue() const { return Val.getInt() == UndefVal; }
570 Value *getSimpleValue() const {
571 assert(isSimpleValue() && "Wrong accessor");
572 return Val.getPointer();
575 LoadInst *getCoercedLoadValue() const {
576 assert(isCoercedLoadValue() && "Wrong accessor");
577 return cast<LoadInst>(Val.getPointer());
580 MemIntrinsic *getMemIntrinValue() const {
581 assert(isMemIntrinValue() && "Wrong accessor");
582 return cast<MemIntrinsic>(Val.getPointer());
585 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
586 /// defined here to the specified type. This handles various coercion cases.
587 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const;
590 class GVN : public FunctionPass {
592 MemoryDependenceAnalysis *MD;
594 const DataLayout *DL;
595 const TargetLibraryInfo *TLI;
597 SetVector<BasicBlock *> DeadBlocks;
601 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
602 /// have that value number. Use findLeader to query it.
603 struct LeaderTableEntry {
605 const BasicBlock *BB;
606 LeaderTableEntry *Next;
608 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
609 BumpPtrAllocator TableAllocator;
611 SmallVector<Instruction*, 8> InstrsToErase;
613 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
614 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
615 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
618 static char ID; // Pass identification, replacement for typeid
619 explicit GVN(bool noloads = false)
620 : FunctionPass(ID), NoLoads(noloads), MD(nullptr) {
621 initializeGVNPass(*PassRegistry::getPassRegistry());
624 bool runOnFunction(Function &F) override;
626 /// markInstructionForDeletion - This removes the specified instruction from
627 /// our various maps and marks it for deletion.
628 void markInstructionForDeletion(Instruction *I) {
630 InstrsToErase.push_back(I);
633 const DataLayout *getDataLayout() const { return DL; }
634 DominatorTree &getDominatorTree() const { return *DT; }
635 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
636 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
638 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
639 /// its value number.
640 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
641 LeaderTableEntry &Curr = LeaderTable[N];
648 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
651 Node->Next = Curr.Next;
655 /// removeFromLeaderTable - Scan the list of values corresponding to a given
656 /// value number, and remove the given instruction if encountered.
657 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
658 LeaderTableEntry* Prev = nullptr;
659 LeaderTableEntry* Curr = &LeaderTable[N];
661 while (Curr->Val != I || Curr->BB != BB) {
667 Prev->Next = Curr->Next;
673 LeaderTableEntry* Next = Curr->Next;
674 Curr->Val = Next->Val;
676 Curr->Next = Next->Next;
681 // List of critical edges to be split between iterations.
682 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
684 // This transformation requires dominator postdominator info
685 void getAnalysisUsage(AnalysisUsage &AU) const override {
686 AU.addRequired<AssumptionCacheTracker>();
687 AU.addRequired<DominatorTreeWrapperPass>();
688 AU.addRequired<TargetLibraryInfoWrapperPass>();
690 AU.addRequired<MemoryDependenceAnalysis>();
691 AU.addRequired<AliasAnalysis>();
693 AU.addPreserved<DominatorTreeWrapperPass>();
694 AU.addPreserved<AliasAnalysis>();
698 // Helper fuctions of redundant load elimination
699 bool processLoad(LoadInst *L);
700 bool processNonLocalLoad(LoadInst *L);
701 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
702 AvailValInBlkVect &ValuesPerBlock,
703 UnavailBlkVect &UnavailableBlocks);
704 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
705 UnavailBlkVect &UnavailableBlocks);
707 // Other helper routines
708 bool processInstruction(Instruction *I);
709 bool processBlock(BasicBlock *BB);
710 void dump(DenseMap<uint32_t, Value*> &d);
711 bool iterateOnFunction(Function &F);
712 bool performPRE(Function &F);
713 bool performScalarPRE(Instruction *I);
714 bool performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
716 Value *findLeader(const BasicBlock *BB, uint32_t num);
717 void cleanupGlobalSets();
718 void verifyRemoved(const Instruction *I) const;
719 bool splitCriticalEdges();
720 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
721 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
722 const BasicBlockEdge &Root);
723 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
724 bool processFoldableCondBr(BranchInst *BI);
725 void addDeadBlock(BasicBlock *BB);
726 void assignValNumForDeadCode();
732 // createGVNPass - The public interface to this file...
733 FunctionPass *llvm::createGVNPass(bool NoLoads) {
734 return new GVN(NoLoads);
737 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
738 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
739 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
740 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
741 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
742 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
743 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
745 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
746 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
748 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
749 E = d.end(); I != E; ++I) {
750 errs() << I->first << "\n";
757 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
758 /// we're analyzing is fully available in the specified block. As we go, keep
759 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
760 /// map is actually a tri-state map with the following values:
761 /// 0) we know the block *is not* fully available.
762 /// 1) we know the block *is* fully available.
763 /// 2) we do not know whether the block is fully available or not, but we are
764 /// currently speculating that it will be.
765 /// 3) we are speculating for this block and have used that to speculate for
767 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
768 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
769 uint32_t RecurseDepth) {
770 if (RecurseDepth > MaxRecurseDepth)
773 // Optimistically assume that the block is fully available and check to see
774 // if we already know about this block in one lookup.
775 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
776 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
778 // If the entry already existed for this block, return the precomputed value.
780 // If this is a speculative "available" value, mark it as being used for
781 // speculation of other blocks.
782 if (IV.first->second == 2)
783 IV.first->second = 3;
784 return IV.first->second != 0;
787 // Otherwise, see if it is fully available in all predecessors.
788 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
790 // If this block has no predecessors, it isn't live-in here.
792 goto SpeculationFailure;
794 for (; PI != PE; ++PI)
795 // If the value isn't fully available in one of our predecessors, then it
796 // isn't fully available in this block either. Undo our previous
797 // optimistic assumption and bail out.
798 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
799 goto SpeculationFailure;
803 // SpeculationFailure - If we get here, we found out that this is not, after
804 // all, a fully-available block. We have a problem if we speculated on this and
805 // used the speculation to mark other blocks as available.
807 char &BBVal = FullyAvailableBlocks[BB];
809 // If we didn't speculate on this, just return with it set to false.
815 // If we did speculate on this value, we could have blocks set to 1 that are
816 // incorrect. Walk the (transitive) successors of this block and mark them as
818 SmallVector<BasicBlock*, 32> BBWorklist;
819 BBWorklist.push_back(BB);
822 BasicBlock *Entry = BBWorklist.pop_back_val();
823 // Note that this sets blocks to 0 (unavailable) if they happen to not
824 // already be in FullyAvailableBlocks. This is safe.
825 char &EntryVal = FullyAvailableBlocks[Entry];
826 if (EntryVal == 0) continue; // Already unavailable.
828 // Mark as unavailable.
831 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
832 } while (!BBWorklist.empty());
838 /// CanCoerceMustAliasedValueToLoad - Return true if
839 /// CoerceAvailableValueToLoadType will succeed.
840 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
842 const DataLayout &DL) {
843 // If the loaded or stored value is an first class array or struct, don't try
844 // to transform them. We need to be able to bitcast to integer.
845 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
846 StoredVal->getType()->isStructTy() ||
847 StoredVal->getType()->isArrayTy())
850 // The store has to be at least as big as the load.
851 if (DL.getTypeSizeInBits(StoredVal->getType()) <
852 DL.getTypeSizeInBits(LoadTy))
858 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
859 /// then a load from a must-aliased pointer of a different type, try to coerce
860 /// the stored value. LoadedTy is the type of the load we want to replace and
861 /// InsertPt is the place to insert new instructions.
863 /// If we can't do it, return null.
864 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
866 Instruction *InsertPt,
867 const DataLayout &DL) {
868 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
871 // If this is already the right type, just return it.
872 Type *StoredValTy = StoredVal->getType();
874 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
875 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
877 // If the store and reload are the same size, we can always reuse it.
878 if (StoreSize == LoadSize) {
879 // Pointer to Pointer -> use bitcast.
880 if (StoredValTy->getScalarType()->isPointerTy() &&
881 LoadedTy->getScalarType()->isPointerTy())
882 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
884 // Convert source pointers to integers, which can be bitcast.
885 if (StoredValTy->getScalarType()->isPointerTy()) {
886 StoredValTy = DL.getIntPtrType(StoredValTy);
887 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
890 Type *TypeToCastTo = LoadedTy;
891 if (TypeToCastTo->getScalarType()->isPointerTy())
892 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
894 if (StoredValTy != TypeToCastTo)
895 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
897 // Cast to pointer if the load needs a pointer type.
898 if (LoadedTy->getScalarType()->isPointerTy())
899 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
904 // If the loaded value is smaller than the available value, then we can
905 // extract out a piece from it. If the available value is too small, then we
906 // can't do anything.
907 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
909 // Convert source pointers to integers, which can be manipulated.
910 if (StoredValTy->getScalarType()->isPointerTy()) {
911 StoredValTy = DL.getIntPtrType(StoredValTy);
912 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
915 // Convert vectors and fp to integer, which can be manipulated.
916 if (!StoredValTy->isIntegerTy()) {
917 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
918 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
921 // If this is a big-endian system, we need to shift the value down to the low
922 // bits so that a truncate will work.
923 if (DL.isBigEndian()) {
924 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
925 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
928 // Truncate the integer to the right size now.
929 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
930 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
932 if (LoadedTy == NewIntTy)
935 // If the result is a pointer, inttoptr.
936 if (LoadedTy->getScalarType()->isPointerTy())
937 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
939 // Otherwise, bitcast.
940 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
943 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
944 /// memdep query of a load that ends up being a clobbering memory write (store,
945 /// memset, memcpy, memmove). This means that the write *may* provide bits used
946 /// by the load but we can't be sure because the pointers don't mustalias.
948 /// Check this case to see if there is anything more we can do before we give
949 /// up. This returns -1 if we have to give up, or a byte number in the stored
950 /// value of the piece that feeds the load.
951 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
953 uint64_t WriteSizeInBits,
954 const DataLayout &DL) {
955 // If the loaded or stored value is a first class array or struct, don't try
956 // to transform them. We need to be able to bitcast to integer.
957 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
960 int64_t StoreOffset = 0, LoadOffset = 0;
961 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&DL);
962 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &DL);
963 if (StoreBase != LoadBase)
966 // If the load and store are to the exact same address, they should have been
967 // a must alias. AA must have gotten confused.
968 // FIXME: Study to see if/when this happens. One case is forwarding a memset
969 // to a load from the base of the memset.
971 if (LoadOffset == StoreOffset) {
972 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
973 << "Base = " << *StoreBase << "\n"
974 << "Store Ptr = " << *WritePtr << "\n"
975 << "Store Offs = " << StoreOffset << "\n"
976 << "Load Ptr = " << *LoadPtr << "\n";
981 // If the load and store don't overlap at all, the store doesn't provide
982 // anything to the load. In this case, they really don't alias at all, AA
983 // must have gotten confused.
984 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
986 if ((WriteSizeInBits & 7) | (LoadSize & 7))
988 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
992 bool isAAFailure = false;
993 if (StoreOffset < LoadOffset)
994 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
996 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1000 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1001 << "Base = " << *StoreBase << "\n"
1002 << "Store Ptr = " << *WritePtr << "\n"
1003 << "Store Offs = " << StoreOffset << "\n"
1004 << "Load Ptr = " << *LoadPtr << "\n";
1010 // If the Load isn't completely contained within the stored bits, we don't
1011 // have all the bits to feed it. We could do something crazy in the future
1012 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1014 if (StoreOffset > LoadOffset ||
1015 StoreOffset+StoreSize < LoadOffset+LoadSize)
1018 // Okay, we can do this transformation. Return the number of bytes into the
1019 // store that the load is.
1020 return LoadOffset-StoreOffset;
1023 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1024 /// memdep query of a load that ends up being a clobbering store.
1025 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1027 const DataLayout &DL) {
1028 // Cannot handle reading from store of first-class aggregate yet.
1029 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1030 DepSI->getValueOperand()->getType()->isArrayTy())
1033 Value *StorePtr = DepSI->getPointerOperand();
1034 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1035 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1036 StorePtr, StoreSize, DL);
1039 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
1040 /// memdep query of a load that ends up being clobbered by another load. See if
1041 /// the other load can feed into the second load.
1042 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1043 LoadInst *DepLI, const DataLayout &DL){
1044 // Cannot handle reading from store of first-class aggregate yet.
1045 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1048 Value *DepPtr = DepLI->getPointerOperand();
1049 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
1050 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
1051 if (R != -1) return R;
1053 // If we have a load/load clobber an DepLI can be widened to cover this load,
1054 // then we should widen it!
1055 int64_t LoadOffs = 0;
1056 const Value *LoadBase =
1057 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &DL);
1058 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1060 unsigned Size = MemoryDependenceAnalysis::
1061 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, DL);
1062 if (Size == 0) return -1;
1064 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
1069 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1071 const DataLayout &DL) {
1072 // If the mem operation is a non-constant size, we can't handle it.
1073 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1074 if (!SizeCst) return -1;
1075 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1077 // If this is memset, we just need to see if the offset is valid in the size
1079 if (MI->getIntrinsicID() == Intrinsic::memset)
1080 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1083 // If we have a memcpy/memmove, the only case we can handle is if this is a
1084 // copy from constant memory. In that case, we can read directly from the
1086 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1088 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1089 if (!Src) return -1;
1091 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &DL));
1092 if (!GV || !GV->isConstant()) return -1;
1094 // See if the access is within the bounds of the transfer.
1095 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1096 MI->getDest(), MemSizeInBits, DL);
1100 unsigned AS = Src->getType()->getPointerAddressSpace();
1101 // Otherwise, see if we can constant fold a load from the constant with the
1102 // offset applied as appropriate.
1103 Src = ConstantExpr::getBitCast(Src,
1104 Type::getInt8PtrTy(Src->getContext(), AS));
1105 Constant *OffsetCst =
1106 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1107 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1108 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1109 if (ConstantFoldLoadFromConstPtr(Src, &DL))
1115 /// GetStoreValueForLoad - This function is called when we have a
1116 /// memdep query of a load that ends up being a clobbering store. This means
1117 /// that the store provides bits used by the load but we the pointers don't
1118 /// mustalias. Check this case to see if there is anything more we can do
1119 /// before we give up.
1120 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1122 Instruction *InsertPt, const DataLayout &DL){
1123 LLVMContext &Ctx = SrcVal->getType()->getContext();
1125 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1126 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1128 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1130 // Compute which bits of the stored value are being used by the load. Convert
1131 // to an integer type to start with.
1132 if (SrcVal->getType()->getScalarType()->isPointerTy())
1133 SrcVal = Builder.CreatePtrToInt(SrcVal,
1134 DL.getIntPtrType(SrcVal->getType()));
1135 if (!SrcVal->getType()->isIntegerTy())
1136 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1138 // Shift the bits to the least significant depending on endianness.
1140 if (DL.isLittleEndian())
1141 ShiftAmt = Offset*8;
1143 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1146 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1148 if (LoadSize != StoreSize)
1149 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1151 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, DL);
1154 /// GetLoadValueForLoad - This function is called when we have a
1155 /// memdep query of a load that ends up being a clobbering load. This means
1156 /// that the load *may* provide bits used by the load but we can't be sure
1157 /// because the pointers don't mustalias. Check this case to see if there is
1158 /// anything more we can do before we give up.
1159 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1160 Type *LoadTy, Instruction *InsertPt,
1162 const DataLayout &DL = *gvn.getDataLayout();
1163 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1164 // widen SrcVal out to a larger load.
1165 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
1166 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1167 if (Offset+LoadSize > SrcValSize) {
1168 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1169 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1170 // If we have a load/load clobber an DepLI can be widened to cover this
1171 // load, then we should widen it to the next power of 2 size big enough!
1172 unsigned NewLoadSize = Offset+LoadSize;
1173 if (!isPowerOf2_32(NewLoadSize))
1174 NewLoadSize = NextPowerOf2(NewLoadSize);
1176 Value *PtrVal = SrcVal->getPointerOperand();
1178 // Insert the new load after the old load. This ensures that subsequent
1179 // memdep queries will find the new load. We can't easily remove the old
1180 // load completely because it is already in the value numbering table.
1181 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1183 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1184 DestPTy = PointerType::get(DestPTy,
1185 PtrVal->getType()->getPointerAddressSpace());
1186 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1187 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1188 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1189 NewLoad->takeName(SrcVal);
1190 NewLoad->setAlignment(SrcVal->getAlignment());
1192 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1193 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1195 // Replace uses of the original load with the wider load. On a big endian
1196 // system, we need to shift down to get the relevant bits.
1197 Value *RV = NewLoad;
1198 if (DL.isBigEndian())
1199 RV = Builder.CreateLShr(RV,
1200 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1201 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1202 SrcVal->replaceAllUsesWith(RV);
1204 // We would like to use gvn.markInstructionForDeletion here, but we can't
1205 // because the load is already memoized into the leader map table that GVN
1206 // tracks. It is potentially possible to remove the load from the table,
1207 // but then there all of the operations based on it would need to be
1208 // rehashed. Just leave the dead load around.
1209 gvn.getMemDep().removeInstruction(SrcVal);
1213 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1217 /// GetMemInstValueForLoad - This function is called when we have a
1218 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1219 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1220 Type *LoadTy, Instruction *InsertPt,
1221 const DataLayout &DL){
1222 LLVMContext &Ctx = LoadTy->getContext();
1223 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1225 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1227 // We know that this method is only called when the mem transfer fully
1228 // provides the bits for the load.
1229 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1230 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1231 // independently of what the offset is.
1232 Value *Val = MSI->getValue();
1234 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1236 Value *OneElt = Val;
1238 // Splat the value out to the right number of bits.
1239 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1240 // If we can double the number of bytes set, do it.
1241 if (NumBytesSet*2 <= LoadSize) {
1242 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1243 Val = Builder.CreateOr(Val, ShVal);
1248 // Otherwise insert one byte at a time.
1249 Value *ShVal = Builder.CreateShl(Val, 1*8);
1250 Val = Builder.CreateOr(OneElt, ShVal);
1254 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, DL);
1257 // Otherwise, this is a memcpy/memmove from a constant global.
1258 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1259 Constant *Src = cast<Constant>(MTI->getSource());
1260 unsigned AS = Src->getType()->getPointerAddressSpace();
1262 // Otherwise, see if we can constant fold a load from the constant with the
1263 // offset applied as appropriate.
1264 Src = ConstantExpr::getBitCast(Src,
1265 Type::getInt8PtrTy(Src->getContext(), AS));
1266 Constant *OffsetCst =
1267 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1268 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1269 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1270 return ConstantFoldLoadFromConstPtr(Src, &DL);
1274 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1275 /// construct SSA form, allowing us to eliminate LI. This returns the value
1276 /// that should be used at LI's definition site.
1277 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1278 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1280 // Check for the fully redundant, dominating load case. In this case, we can
1281 // just use the dominating value directly.
1282 if (ValuesPerBlock.size() == 1 &&
1283 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1285 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1286 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1289 // Otherwise, we have to construct SSA form.
1290 SmallVector<PHINode*, 8> NewPHIs;
1291 SSAUpdater SSAUpdate(&NewPHIs);
1292 SSAUpdate.Initialize(LI->getType(), LI->getName());
1294 Type *LoadTy = LI->getType();
1296 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1297 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1298 BasicBlock *BB = AV.BB;
1300 if (SSAUpdate.HasValueForBlock(BB))
1303 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1306 // Perform PHI construction.
1307 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1309 // If new PHI nodes were created, notify alias analysis.
1310 if (V->getType()->getScalarType()->isPointerTy()) {
1311 AliasAnalysis *AA = gvn.getAliasAnalysis();
1313 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1314 AA->copyValue(LI, NewPHIs[i]);
1316 // Now that we've copied information to the new PHIs, scan through
1317 // them again and inform alias analysis that we've added potentially
1318 // escaping uses to any values that are operands to these PHIs.
1319 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1320 PHINode *P = NewPHIs[i];
1321 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1322 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1323 AA->addEscapingUse(P->getOperandUse(jj));
1331 Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1333 if (isSimpleValue()) {
1334 Res = getSimpleValue();
1335 if (Res->getType() != LoadTy) {
1336 const DataLayout *DL = gvn.getDataLayout();
1337 assert(DL && "Need target data to handle type mismatch case");
1338 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1341 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1342 << *getSimpleValue() << '\n'
1343 << *Res << '\n' << "\n\n\n");
1345 } else if (isCoercedLoadValue()) {
1346 LoadInst *Load = getCoercedLoadValue();
1347 if (Load->getType() == LoadTy && Offset == 0) {
1350 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1353 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1354 << *getCoercedLoadValue() << '\n'
1355 << *Res << '\n' << "\n\n\n");
1357 } else if (isMemIntrinValue()) {
1358 const DataLayout *DL = gvn.getDataLayout();
1359 assert(DL && "Need target data to handle type mismatch case");
1360 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1361 LoadTy, BB->getTerminator(), *DL);
1362 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1363 << " " << *getMemIntrinValue() << '\n'
1364 << *Res << '\n' << "\n\n\n");
1366 assert(isUndefValue() && "Should be UndefVal");
1367 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1368 return UndefValue::get(LoadTy);
1373 static bool isLifetimeStart(const Instruction *Inst) {
1374 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1375 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1379 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1380 AvailValInBlkVect &ValuesPerBlock,
1381 UnavailBlkVect &UnavailableBlocks) {
1383 // Filter out useless results (non-locals, etc). Keep track of the blocks
1384 // where we have a value available in repl, also keep track of whether we see
1385 // dependencies that produce an unknown value for the load (such as a call
1386 // that could potentially clobber the load).
1387 unsigned NumDeps = Deps.size();
1388 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1389 BasicBlock *DepBB = Deps[i].getBB();
1390 MemDepResult DepInfo = Deps[i].getResult();
1392 if (DeadBlocks.count(DepBB)) {
1393 // Dead dependent mem-op disguise as a load evaluating the same value
1394 // as the load in question.
1395 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1399 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1400 UnavailableBlocks.push_back(DepBB);
1404 if (DepInfo.isClobber()) {
1405 // The address being loaded in this non-local block may not be the same as
1406 // the pointer operand of the load if PHI translation occurs. Make sure
1407 // to consider the right address.
1408 Value *Address = Deps[i].getAddress();
1410 // If the dependence is to a store that writes to a superset of the bits
1411 // read by the load, we can extract the bits we need for the load from the
1413 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1414 if (DL && Address) {
1415 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1418 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1419 DepSI->getValueOperand(),
1426 // Check to see if we have something like this:
1429 // if we have this, replace the later with an extraction from the former.
1430 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1431 // If this is a clobber and L is the first instruction in its block, then
1432 // we have the first instruction in the entry block.
1433 if (DepLI != LI && Address && DL) {
1434 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), Address,
1438 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1445 // If the clobbering value is a memset/memcpy/memmove, see if we can
1446 // forward a value on from it.
1447 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1448 if (DL && Address) {
1449 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1452 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1459 UnavailableBlocks.push_back(DepBB);
1463 // DepInfo.isDef() here
1465 Instruction *DepInst = DepInfo.getInst();
1467 // Loading the allocation -> undef.
1468 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1469 // Loading immediately after lifetime begin -> undef.
1470 isLifetimeStart(DepInst)) {
1471 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1472 UndefValue::get(LI->getType())));
1476 // Loading from calloc (which zero initializes memory) -> zero
1477 if (isCallocLikeFn(DepInst, TLI)) {
1478 ValuesPerBlock.push_back(AvailableValueInBlock::get(
1479 DepBB, Constant::getNullValue(LI->getType())));
1483 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1484 // Reject loads and stores that are to the same address but are of
1485 // different types if we have to.
1486 if (S->getValueOperand()->getType() != LI->getType()) {
1487 // If the stored value is larger or equal to the loaded value, we can
1489 if (!DL || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1490 LI->getType(), *DL)) {
1491 UnavailableBlocks.push_back(DepBB);
1496 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1497 S->getValueOperand()));
1501 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1502 // If the types mismatch and we can't handle it, reject reuse of the load.
1503 if (LD->getType() != LI->getType()) {
1504 // If the stored value is larger or equal to the loaded value, we can
1506 if (!DL || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*DL)) {
1507 UnavailableBlocks.push_back(DepBB);
1511 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1515 UnavailableBlocks.push_back(DepBB);
1519 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1520 UnavailBlkVect &UnavailableBlocks) {
1521 // Okay, we have *some* definitions of the value. This means that the value
1522 // is available in some of our (transitive) predecessors. Lets think about
1523 // doing PRE of this load. This will involve inserting a new load into the
1524 // predecessor when it's not available. We could do this in general, but
1525 // prefer to not increase code size. As such, we only do this when we know
1526 // that we only have to insert *one* load (which means we're basically moving
1527 // the load, not inserting a new one).
1529 SmallPtrSet<BasicBlock *, 4> Blockers;
1530 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1531 Blockers.insert(UnavailableBlocks[i]);
1533 // Let's find the first basic block with more than one predecessor. Walk
1534 // backwards through predecessors if needed.
1535 BasicBlock *LoadBB = LI->getParent();
1536 BasicBlock *TmpBB = LoadBB;
1538 while (TmpBB->getSinglePredecessor()) {
1539 TmpBB = TmpBB->getSinglePredecessor();
1540 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1542 if (Blockers.count(TmpBB))
1545 // If any of these blocks has more than one successor (i.e. if the edge we
1546 // just traversed was critical), then there are other paths through this
1547 // block along which the load may not be anticipated. Hoisting the load
1548 // above this block would be adding the load to execution paths along
1549 // which it was not previously executed.
1550 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1557 // Check to see how many predecessors have the loaded value fully
1559 MapVector<BasicBlock *, Value *> PredLoads;
1560 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1561 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1562 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1563 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1564 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1566 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1567 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1569 BasicBlock *Pred = *PI;
1570 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1574 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1575 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1576 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1577 << Pred->getName() << "': " << *LI << '\n');
1581 if (LoadBB->isLandingPad()) {
1583 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1584 << Pred->getName() << "': " << *LI << '\n');
1588 CriticalEdgePred.push_back(Pred);
1590 // Only add the predecessors that will not be split for now.
1591 PredLoads[Pred] = nullptr;
1595 // Decide whether PRE is profitable for this load.
1596 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1597 assert(NumUnavailablePreds != 0 &&
1598 "Fully available value should already be eliminated!");
1600 // If this load is unavailable in multiple predecessors, reject it.
1601 // FIXME: If we could restructure the CFG, we could make a common pred with
1602 // all the preds that don't have an available LI and insert a new load into
1604 if (NumUnavailablePreds != 1)
1607 // Split critical edges, and update the unavailable predecessors accordingly.
1608 for (BasicBlock *OrigPred : CriticalEdgePred) {
1609 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1610 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1611 PredLoads[NewPred] = nullptr;
1612 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1613 << LoadBB->getName() << '\n');
1616 // Check if the load can safely be moved to all the unavailable predecessors.
1617 bool CanDoPRE = true;
1618 SmallVector<Instruction*, 8> NewInsts;
1619 for (auto &PredLoad : PredLoads) {
1620 BasicBlock *UnavailablePred = PredLoad.first;
1622 // Do PHI translation to get its value in the predecessor if necessary. The
1623 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1625 // If all preds have a single successor, then we know it is safe to insert
1626 // the load on the pred (?!?), so we can insert code to materialize the
1627 // pointer if it is not available.
1628 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1629 Value *LoadPtr = nullptr;
1630 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1633 // If we couldn't find or insert a computation of this phi translated value,
1636 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1637 << *LI->getPointerOperand() << "\n");
1642 PredLoad.second = LoadPtr;
1646 while (!NewInsts.empty()) {
1647 Instruction *I = NewInsts.pop_back_val();
1648 if (MD) MD->removeInstruction(I);
1649 I->eraseFromParent();
1651 // HINT: Don't revert the edge-splitting as following transformation may
1652 // also need to split these critical edges.
1653 return !CriticalEdgePred.empty();
1656 // Okay, we can eliminate this load by inserting a reload in the predecessor
1657 // and using PHI construction to get the value in the other predecessors, do
1659 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1660 DEBUG(if (!NewInsts.empty())
1661 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1662 << *NewInsts.back() << '\n');
1664 // Assign value numbers to the new instructions.
1665 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1666 // FIXME: We really _ought_ to insert these value numbers into their
1667 // parent's availability map. However, in doing so, we risk getting into
1668 // ordering issues. If a block hasn't been processed yet, we would be
1669 // marking a value as AVAIL-IN, which isn't what we intend.
1670 VN.lookup_or_add(NewInsts[i]);
1673 for (const auto &PredLoad : PredLoads) {
1674 BasicBlock *UnavailablePred = PredLoad.first;
1675 Value *LoadPtr = PredLoad.second;
1677 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1679 UnavailablePred->getTerminator());
1681 // Transfer the old load's AA tags to the new load.
1683 LI->getAAMetadata(Tags);
1685 NewLoad->setAAMetadata(Tags);
1687 // Transfer DebugLoc.
1688 NewLoad->setDebugLoc(LI->getDebugLoc());
1690 // Add the newly created load.
1691 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1693 MD->invalidateCachedPointerInfo(LoadPtr);
1694 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1697 // Perform PHI construction.
1698 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1699 LI->replaceAllUsesWith(V);
1700 if (isa<PHINode>(V))
1702 if (V->getType()->getScalarType()->isPointerTy())
1703 MD->invalidateCachedPointerInfo(V);
1704 markInstructionForDeletion(LI);
1709 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1710 /// non-local by performing PHI construction.
1711 bool GVN::processNonLocalLoad(LoadInst *LI) {
1712 // Step 1: Find the non-local dependencies of the load.
1714 MD->getNonLocalPointerDependency(LI, Deps);
1716 // If we had to process more than one hundred blocks to find the
1717 // dependencies, this load isn't worth worrying about. Optimizing
1718 // it will be too expensive.
1719 unsigned NumDeps = Deps.size();
1723 // If we had a phi translation failure, we'll have a single entry which is a
1724 // clobber in the current block. Reject this early.
1726 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1728 dbgs() << "GVN: non-local load ";
1729 LI->printAsOperand(dbgs());
1730 dbgs() << " has unknown dependencies\n";
1735 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1736 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1737 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1738 OE = GEP->idx_end();
1740 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1741 performScalarPRE(I);
1744 // Step 2: Analyze the availability of the load
1745 AvailValInBlkVect ValuesPerBlock;
1746 UnavailBlkVect UnavailableBlocks;
1747 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1749 // If we have no predecessors that produce a known value for this load, exit
1751 if (ValuesPerBlock.empty())
1754 // Step 3: Eliminate fully redundancy.
1756 // If all of the instructions we depend on produce a known value for this
1757 // load, then it is fully redundant and we can use PHI insertion to compute
1758 // its value. Insert PHIs and remove the fully redundant value now.
1759 if (UnavailableBlocks.empty()) {
1760 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1762 // Perform PHI construction.
1763 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1764 LI->replaceAllUsesWith(V);
1766 if (isa<PHINode>(V))
1768 if (V->getType()->getScalarType()->isPointerTy())
1769 MD->invalidateCachedPointerInfo(V);
1770 markInstructionForDeletion(LI);
1775 // Step 4: Eliminate partial redundancy.
1776 if (!EnablePRE || !EnableLoadPRE)
1779 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1783 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1784 // Patch the replacement so that it is not more restrictive than the value
1786 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1787 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1788 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
1789 isa<OverflowingBinaryOperator>(ReplOp)) {
1790 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
1791 ReplOp->setHasNoSignedWrap(false);
1792 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
1793 ReplOp->setHasNoUnsignedWrap(false);
1795 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1796 // FIXME: If both the original and replacement value are part of the
1797 // same control-flow region (meaning that the execution of one
1798 // guarentees the executation of the other), then we can combine the
1799 // noalias scopes here and do better than the general conservative
1800 // answer used in combineMetadata().
1802 // In general, GVN unifies expressions over different control-flow
1803 // regions, and so we need a conservative combination of the noalias
1805 unsigned KnownIDs[] = {
1806 LLVMContext::MD_tbaa,
1807 LLVMContext::MD_alias_scope,
1808 LLVMContext::MD_noalias,
1809 LLVMContext::MD_range,
1810 LLVMContext::MD_fpmath,
1811 LLVMContext::MD_invariant_load,
1813 combineMetadata(ReplInst, I, KnownIDs);
1817 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1818 patchReplacementInstruction(I, Repl);
1819 I->replaceAllUsesWith(Repl);
1822 /// processLoad - Attempt to eliminate a load, first by eliminating it
1823 /// locally, and then attempting non-local elimination if that fails.
1824 bool GVN::processLoad(LoadInst *L) {
1831 if (L->use_empty()) {
1832 markInstructionForDeletion(L);
1836 // ... to a pointer that has been loaded from before...
1837 MemDepResult Dep = MD->getDependency(L);
1839 // If we have a clobber and target data is around, see if this is a clobber
1840 // that we can fix up through code synthesis.
1841 if (Dep.isClobber() && DL) {
1842 // Check to see if we have something like this:
1843 // store i32 123, i32* %P
1844 // %A = bitcast i32* %P to i8*
1845 // %B = gep i8* %A, i32 1
1848 // We could do that by recognizing if the clobber instructions are obviously
1849 // a common base + constant offset, and if the previous store (or memset)
1850 // completely covers this load. This sort of thing can happen in bitfield
1852 Value *AvailVal = nullptr;
1853 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1854 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1855 L->getPointerOperand(),
1858 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1859 L->getType(), L, *DL);
1862 // Check to see if we have something like this:
1865 // if we have this, replace the later with an extraction from the former.
1866 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1867 // If this is a clobber and L is the first instruction in its block, then
1868 // we have the first instruction in the entry block.
1872 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1873 L->getPointerOperand(),
1876 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1879 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1880 // a value on from it.
1881 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1882 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1883 L->getPointerOperand(),
1886 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *DL);
1890 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1891 << *AvailVal << '\n' << *L << "\n\n\n");
1893 // Replace the load!
1894 L->replaceAllUsesWith(AvailVal);
1895 if (AvailVal->getType()->getScalarType()->isPointerTy())
1896 MD->invalidateCachedPointerInfo(AvailVal);
1897 markInstructionForDeletion(L);
1903 // If the value isn't available, don't do anything!
1904 if (Dep.isClobber()) {
1906 // fast print dep, using operator<< on instruction is too slow.
1907 dbgs() << "GVN: load ";
1908 L->printAsOperand(dbgs());
1909 Instruction *I = Dep.getInst();
1910 dbgs() << " is clobbered by " << *I << '\n';
1915 // If it is defined in another block, try harder.
1916 if (Dep.isNonLocal())
1917 return processNonLocalLoad(L);
1921 // fast print dep, using operator<< on instruction is too slow.
1922 dbgs() << "GVN: load ";
1923 L->printAsOperand(dbgs());
1924 dbgs() << " has unknown dependence\n";
1929 Instruction *DepInst = Dep.getInst();
1930 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1931 Value *StoredVal = DepSI->getValueOperand();
1933 // The store and load are to a must-aliased pointer, but they may not
1934 // actually have the same type. See if we know how to reuse the stored
1935 // value (depending on its type).
1936 if (StoredVal->getType() != L->getType()) {
1938 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1943 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1944 << '\n' << *L << "\n\n\n");
1951 L->replaceAllUsesWith(StoredVal);
1952 if (StoredVal->getType()->getScalarType()->isPointerTy())
1953 MD->invalidateCachedPointerInfo(StoredVal);
1954 markInstructionForDeletion(L);
1959 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1960 Value *AvailableVal = DepLI;
1962 // The loads are of a must-aliased pointer, but they may not actually have
1963 // the same type. See if we know how to reuse the previously loaded value
1964 // (depending on its type).
1965 if (DepLI->getType() != L->getType()) {
1967 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1972 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1973 << "\n" << *L << "\n\n\n");
1980 patchAndReplaceAllUsesWith(L, AvailableVal);
1981 if (DepLI->getType()->getScalarType()->isPointerTy())
1982 MD->invalidateCachedPointerInfo(DepLI);
1983 markInstructionForDeletion(L);
1988 // If this load really doesn't depend on anything, then we must be loading an
1989 // undef value. This can happen when loading for a fresh allocation with no
1990 // intervening stores, for example.
1991 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1992 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1993 markInstructionForDeletion(L);
1998 // If this load occurs either right after a lifetime begin,
1999 // then the loaded value is undefined.
2000 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
2001 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
2002 L->replaceAllUsesWith(UndefValue::get(L->getType()));
2003 markInstructionForDeletion(L);
2009 // If this load follows a calloc (which zero initializes memory),
2010 // then the loaded value is zero
2011 if (isCallocLikeFn(DepInst, TLI)) {
2012 L->replaceAllUsesWith(Constant::getNullValue(L->getType()));
2013 markInstructionForDeletion(L);
2021 // findLeader - In order to find a leader for a given value number at a
2022 // specific basic block, we first obtain the list of all Values for that number,
2023 // and then scan the list to find one whose block dominates the block in
2024 // question. This is fast because dominator tree queries consist of only
2025 // a few comparisons of DFS numbers.
2026 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2027 LeaderTableEntry Vals = LeaderTable[num];
2028 if (!Vals.Val) return nullptr;
2030 Value *Val = nullptr;
2031 if (DT->dominates(Vals.BB, BB)) {
2033 if (isa<Constant>(Val)) return Val;
2036 LeaderTableEntry* Next = Vals.Next;
2038 if (DT->dominates(Next->BB, BB)) {
2039 if (isa<Constant>(Next->Val)) return Next->Val;
2040 if (!Val) Val = Next->Val;
2049 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
2050 /// use is dominated by the given basic block. Returns the number of uses that
2052 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
2053 const BasicBlockEdge &Root) {
2055 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2059 if (DT->dominates(Root, U)) {
2067 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2068 /// true if every path from the entry block to 'Dst' passes via this edge. In
2069 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2070 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2071 DominatorTree *DT) {
2072 // While in theory it is interesting to consider the case in which Dst has
2073 // more than one predecessor, because Dst might be part of a loop which is
2074 // only reachable from Src, in practice it is pointless since at the time
2075 // GVN runs all such loops have preheaders, which means that Dst will have
2076 // been changed to have only one predecessor, namely Src.
2077 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2078 const BasicBlock *Src = E.getStart();
2079 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2081 return Pred != nullptr;
2084 /// propagateEquality - The given values are known to be equal in every block
2085 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2086 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2087 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2088 const BasicBlockEdge &Root) {
2089 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2090 Worklist.push_back(std::make_pair(LHS, RHS));
2091 bool Changed = false;
2092 // For speed, compute a conservative fast approximation to
2093 // DT->dominates(Root, Root.getEnd());
2094 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2096 while (!Worklist.empty()) {
2097 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2098 LHS = Item.first; RHS = Item.second;
2100 if (LHS == RHS) continue;
2101 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2103 // Don't try to propagate equalities between constants.
2104 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2106 // Prefer a constant on the right-hand side, or an Argument if no constants.
2107 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2108 std::swap(LHS, RHS);
2109 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2111 // If there is no obvious reason to prefer the left-hand side over the
2112 // right-hand side, ensure the longest lived term is on the right-hand side,
2113 // so the shortest lived term will be replaced by the longest lived.
2114 // This tends to expose more simplifications.
2115 uint32_t LVN = VN.lookup_or_add(LHS);
2116 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2117 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2118 // Move the 'oldest' value to the right-hand side, using the value number
2119 // as a proxy for age.
2120 uint32_t RVN = VN.lookup_or_add(RHS);
2122 std::swap(LHS, RHS);
2127 // If value numbering later sees that an instruction in the scope is equal
2128 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2129 // the invariant that instructions only occur in the leader table for their
2130 // own value number (this is used by removeFromLeaderTable), do not do this
2131 // if RHS is an instruction (if an instruction in the scope is morphed into
2132 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2133 // using the leader table is about compiling faster, not optimizing better).
2134 // The leader table only tracks basic blocks, not edges. Only add to if we
2135 // have the simple case where the edge dominates the end.
2136 if (RootDominatesEnd && !isa<Instruction>(RHS))
2137 addToLeaderTable(LVN, RHS, Root.getEnd());
2139 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2140 // LHS always has at least one use that is not dominated by Root, this will
2141 // never do anything if LHS has only one use.
2142 if (!LHS->hasOneUse()) {
2143 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2144 Changed |= NumReplacements > 0;
2145 NumGVNEqProp += NumReplacements;
2148 // Now try to deduce additional equalities from this one. For example, if
2149 // the known equality was "(A != B)" == "false" then it follows that A and B
2150 // are equal in the scope. Only boolean equalities with an explicit true or
2151 // false RHS are currently supported.
2152 if (!RHS->getType()->isIntegerTy(1))
2153 // Not a boolean equality - bail out.
2155 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2157 // RHS neither 'true' nor 'false' - bail out.
2159 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2160 bool isKnownTrue = CI->isAllOnesValue();
2161 bool isKnownFalse = !isKnownTrue;
2163 // If "A && B" is known true then both A and B are known true. If "A || B"
2164 // is known false then both A and B are known false.
2166 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2167 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2168 Worklist.push_back(std::make_pair(A, RHS));
2169 Worklist.push_back(std::make_pair(B, RHS));
2173 // If we are propagating an equality like "(A == B)" == "true" then also
2174 // propagate the equality A == B. When propagating a comparison such as
2175 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2176 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
2177 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2179 // If "A == B" is known true, or "A != B" is known false, then replace
2180 // A with B everywhere in the scope.
2181 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2182 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2183 Worklist.push_back(std::make_pair(Op0, Op1));
2185 // Handle the floating point versions of equality comparisons too.
2186 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
2187 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
2188 // Floating point -0.0 and 0.0 compare equal, so we can't
2189 // propagate a constant based on that comparison.
2190 // FIXME: We should do this optimization if 'no signed zeros' is
2191 // applicable via an instruction-level fast-math-flag or some other
2192 // indicator that relaxed FP semantics are being used.
2193 if (!isa<ConstantFP>(Op1) || !cast<ConstantFP>(Op1)->isZero())
2194 Worklist.push_back(std::make_pair(Op0, Op1));
2197 // If "A >= B" is known true, replace "A < B" with false everywhere.
2198 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2199 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2200 // Since we don't have the instruction "A < B" immediately to hand, work
2201 // out the value number that it would have and use that to find an
2202 // appropriate instruction (if any).
2203 uint32_t NextNum = VN.getNextUnusedValueNumber();
2204 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2205 // If the number we were assigned was brand new then there is no point in
2206 // looking for an instruction realizing it: there cannot be one!
2207 if (Num < NextNum) {
2208 Value *NotCmp = findLeader(Root.getEnd(), Num);
2209 if (NotCmp && isa<Instruction>(NotCmp)) {
2210 unsigned NumReplacements =
2211 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2212 Changed |= NumReplacements > 0;
2213 NumGVNEqProp += NumReplacements;
2216 // Ensure that any instruction in scope that gets the "A < B" value number
2217 // is replaced with false.
2218 // The leader table only tracks basic blocks, not edges. Only add to if we
2219 // have the simple case where the edge dominates the end.
2220 if (RootDominatesEnd)
2221 addToLeaderTable(Num, NotVal, Root.getEnd());
2230 /// processInstruction - When calculating availability, handle an instruction
2231 /// by inserting it into the appropriate sets
2232 bool GVN::processInstruction(Instruction *I) {
2233 // Ignore dbg info intrinsics.
2234 if (isa<DbgInfoIntrinsic>(I))
2237 // If the instruction can be easily simplified then do so now in preference
2238 // to value numbering it. Value numbering often exposes redundancies, for
2239 // example if it determines that %y is equal to %x then the instruction
2240 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2241 if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
2242 I->replaceAllUsesWith(V);
2243 if (MD && V->getType()->getScalarType()->isPointerTy())
2244 MD->invalidateCachedPointerInfo(V);
2245 markInstructionForDeletion(I);
2250 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2251 if (processLoad(LI))
2254 unsigned Num = VN.lookup_or_add(LI);
2255 addToLeaderTable(Num, LI, LI->getParent());
2259 // For conditional branches, we can perform simple conditional propagation on
2260 // the condition value itself.
2261 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2262 if (!BI->isConditional())
2265 if (isa<Constant>(BI->getCondition()))
2266 return processFoldableCondBr(BI);
2268 Value *BranchCond = BI->getCondition();
2269 BasicBlock *TrueSucc = BI->getSuccessor(0);
2270 BasicBlock *FalseSucc = BI->getSuccessor(1);
2271 // Avoid multiple edges early.
2272 if (TrueSucc == FalseSucc)
2275 BasicBlock *Parent = BI->getParent();
2276 bool Changed = false;
2278 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2279 BasicBlockEdge TrueE(Parent, TrueSucc);
2280 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2282 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2283 BasicBlockEdge FalseE(Parent, FalseSucc);
2284 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2289 // For switches, propagate the case values into the case destinations.
2290 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2291 Value *SwitchCond = SI->getCondition();
2292 BasicBlock *Parent = SI->getParent();
2293 bool Changed = false;
2295 // Remember how many outgoing edges there are to every successor.
2296 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2297 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2298 ++SwitchEdges[SI->getSuccessor(i)];
2300 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2302 BasicBlock *Dst = i.getCaseSuccessor();
2303 // If there is only a single edge, propagate the case value into it.
2304 if (SwitchEdges.lookup(Dst) == 1) {
2305 BasicBlockEdge E(Parent, Dst);
2306 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2312 // Instructions with void type don't return a value, so there's
2313 // no point in trying to find redundancies in them.
2314 if (I->getType()->isVoidTy()) return false;
2316 uint32_t NextNum = VN.getNextUnusedValueNumber();
2317 unsigned Num = VN.lookup_or_add(I);
2319 // Allocations are always uniquely numbered, so we can save time and memory
2320 // by fast failing them.
2321 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2322 addToLeaderTable(Num, I, I->getParent());
2326 // If the number we were assigned was a brand new VN, then we don't
2327 // need to do a lookup to see if the number already exists
2328 // somewhere in the domtree: it can't!
2329 if (Num >= NextNum) {
2330 addToLeaderTable(Num, I, I->getParent());
2334 // Perform fast-path value-number based elimination of values inherited from
2336 Value *repl = findLeader(I->getParent(), Num);
2338 // Failure, just remember this instance for future use.
2339 addToLeaderTable(Num, I, I->getParent());
2344 patchAndReplaceAllUsesWith(I, repl);
2345 if (MD && repl->getType()->getScalarType()->isPointerTy())
2346 MD->invalidateCachedPointerInfo(repl);
2347 markInstructionForDeletion(I);
2351 /// runOnFunction - This is the main transformation entry point for a function.
2352 bool GVN::runOnFunction(Function& F) {
2353 if (skipOptnoneFunction(F))
2357 MD = &getAnalysis<MemoryDependenceAnalysis>();
2358 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2359 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2360 DL = DLP ? &DLP->getDataLayout() : nullptr;
2361 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2362 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
2363 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2367 bool Changed = false;
2368 bool ShouldContinue = true;
2370 // Merge unconditional branches, allowing PRE to catch more
2371 // optimization opportunities.
2372 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2373 BasicBlock *BB = FI++;
2375 bool removedBlock = MergeBlockIntoPredecessor(
2376 BB, DT, /* LoopInfo */ nullptr, VN.getAliasAnalysis(), MD);
2377 if (removedBlock) ++NumGVNBlocks;
2379 Changed |= removedBlock;
2382 unsigned Iteration = 0;
2383 while (ShouldContinue) {
2384 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2385 ShouldContinue = iterateOnFunction(F);
2386 Changed |= ShouldContinue;
2391 // Fabricate val-num for dead-code in order to suppress assertion in
2393 assignValNumForDeadCode();
2394 bool PREChanged = true;
2395 while (PREChanged) {
2396 PREChanged = performPRE(F);
2397 Changed |= PREChanged;
2401 // FIXME: Should perform GVN again after PRE does something. PRE can move
2402 // computations into blocks where they become fully redundant. Note that
2403 // we can't do this until PRE's critical edge splitting updates memdep.
2404 // Actually, when this happens, we should just fully integrate PRE into GVN.
2406 cleanupGlobalSets();
2407 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2415 bool GVN::processBlock(BasicBlock *BB) {
2416 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2417 // (and incrementing BI before processing an instruction).
2418 assert(InstrsToErase.empty() &&
2419 "We expect InstrsToErase to be empty across iterations");
2420 if (DeadBlocks.count(BB))
2423 bool ChangedFunction = false;
2425 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2427 ChangedFunction |= processInstruction(BI);
2428 if (InstrsToErase.empty()) {
2433 // If we need some instructions deleted, do it now.
2434 NumGVNInstr += InstrsToErase.size();
2436 // Avoid iterator invalidation.
2437 bool AtStart = BI == BB->begin();
2441 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2442 E = InstrsToErase.end(); I != E; ++I) {
2443 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2444 if (MD) MD->removeInstruction(*I);
2445 DEBUG(verifyRemoved(*I));
2446 (*I)->eraseFromParent();
2448 InstrsToErase.clear();
2456 return ChangedFunction;
2459 // Instantiate an expression in a predecessor that lacked it.
2460 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2461 unsigned int ValNo) {
2462 // Because we are going top-down through the block, all value numbers
2463 // will be available in the predecessor by the time we need them. Any
2464 // that weren't originally present will have been instantiated earlier
2466 bool success = true;
2467 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2468 Value *Op = Instr->getOperand(i);
2469 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2472 if (Value *V = findLeader(Pred, VN.lookup(Op))) {
2473 Instr->setOperand(i, V);
2480 // Fail out if we encounter an operand that is not available in
2481 // the PRE predecessor. This is typically because of loads which
2482 // are not value numbered precisely.
2486 Instr->insertBefore(Pred->getTerminator());
2487 Instr->setName(Instr->getName() + ".pre");
2488 Instr->setDebugLoc(Instr->getDebugLoc());
2489 VN.add(Instr, ValNo);
2491 // Update the availability map to include the new instruction.
2492 addToLeaderTable(ValNo, Instr, Pred);
2496 bool GVN::performScalarPRE(Instruction *CurInst) {
2497 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2499 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2500 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2501 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2502 isa<DbgInfoIntrinsic>(CurInst))
2505 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2506 // sinking the compare again, and it would force the code generator to
2507 // move the i1 from processor flags or predicate registers into a general
2508 // purpose register.
2509 if (isa<CmpInst>(CurInst))
2512 // We don't currently value number ANY inline asm calls.
2513 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2514 if (CallI->isInlineAsm())
2517 uint32_t ValNo = VN.lookup(CurInst);
2519 // Look for the predecessors for PRE opportunities. We're
2520 // only trying to solve the basic diamond case, where
2521 // a value is computed in the successor and one predecessor,
2522 // but not the other. We also explicitly disallow cases
2523 // where the successor is its own predecessor, because they're
2524 // more complicated to get right.
2525 unsigned NumWith = 0;
2526 unsigned NumWithout = 0;
2527 BasicBlock *PREPred = nullptr;
2528 BasicBlock *CurrentBlock = CurInst->getParent();
2531 for (pred_iterator PI = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2533 BasicBlock *P = *PI;
2534 // We're not interested in PRE where the block is its
2535 // own predecessor, or in blocks with predecessors
2536 // that are not reachable.
2537 if (P == CurrentBlock) {
2540 } else if (!DT->isReachableFromEntry(P)) {
2545 Value *predV = findLeader(P, ValNo);
2547 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2550 } else if (predV == CurInst) {
2551 /* CurInst dominates this predecessor. */
2555 predMap.push_back(std::make_pair(predV, P));
2560 // Don't do PRE when it might increase code size, i.e. when
2561 // we would need to insert instructions in more than one pred.
2562 if (NumWithout > 1 || NumWith == 0)
2565 // We may have a case where all predecessors have the instruction,
2566 // and we just need to insert a phi node. Otherwise, perform
2568 Instruction *PREInstr = nullptr;
2570 if (NumWithout != 0) {
2571 // Don't do PRE across indirect branch.
2572 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2575 // We can't do PRE safely on a critical edge, so instead we schedule
2576 // the edge to be split and perform the PRE the next time we iterate
2578 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2579 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2580 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2583 // We need to insert somewhere, so let's give it a shot
2584 PREInstr = CurInst->clone();
2585 if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2586 // If we failed insertion, make sure we remove the instruction.
2587 DEBUG(verifyRemoved(PREInstr));
2593 // Either we should have filled in the PRE instruction, or we should
2594 // not have needed insertions.
2595 assert (PREInstr != nullptr || NumWithout == 0);
2599 // Create a PHI to make the value available in this block.
2601 PHINode::Create(CurInst->getType(), predMap.size(),
2602 CurInst->getName() + ".pre-phi", CurrentBlock->begin());
2603 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2604 if (Value *V = predMap[i].first)
2605 Phi->addIncoming(V, predMap[i].second);
2607 Phi->addIncoming(PREInstr, PREPred);
2611 addToLeaderTable(ValNo, Phi, CurrentBlock);
2612 Phi->setDebugLoc(CurInst->getDebugLoc());
2613 CurInst->replaceAllUsesWith(Phi);
2614 if (Phi->getType()->getScalarType()->isPointerTy()) {
2615 // Because we have added a PHI-use of the pointer value, it has now
2616 // "escaped" from alias analysis' perspective. We need to inform
2618 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii) {
2619 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2620 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2624 MD->invalidateCachedPointerInfo(Phi);
2627 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2629 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2631 MD->removeInstruction(CurInst);
2632 DEBUG(verifyRemoved(CurInst));
2633 CurInst->eraseFromParent();
2639 /// performPRE - Perform a purely local form of PRE that looks for diamond
2640 /// control flow patterns and attempts to perform simple PRE at the join point.
2641 bool GVN::performPRE(Function &F) {
2642 bool Changed = false;
2643 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2644 // Nothing to PRE in the entry block.
2645 if (CurrentBlock == &F.getEntryBlock())
2648 // Don't perform PRE on a landing pad.
2649 if (CurrentBlock->isLandingPad())
2652 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2653 BE = CurrentBlock->end();
2655 Instruction *CurInst = BI++;
2656 Changed = performScalarPRE(CurInst);
2660 if (splitCriticalEdges())
2666 /// Split the critical edge connecting the given two blocks, and return
2667 /// the block inserted to the critical edge.
2668 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2669 BasicBlock *BB = SplitCriticalEdge(
2670 Pred, Succ, CriticalEdgeSplittingOptions(getAliasAnalysis(), DT));
2672 MD->invalidateCachedPredecessors();
2676 /// splitCriticalEdges - Split critical edges found during the previous
2677 /// iteration that may enable further optimization.
2678 bool GVN::splitCriticalEdges() {
2679 if (toSplit.empty())
2682 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2683 SplitCriticalEdge(Edge.first, Edge.second,
2684 CriticalEdgeSplittingOptions(getAliasAnalysis(), DT));
2685 } while (!toSplit.empty());
2686 if (MD) MD->invalidateCachedPredecessors();
2690 /// iterateOnFunction - Executes one iteration of GVN
2691 bool GVN::iterateOnFunction(Function &F) {
2692 cleanupGlobalSets();
2694 // Top-down walk of the dominator tree
2695 bool Changed = false;
2696 // Save the blocks this function have before transformation begins. GVN may
2697 // split critical edge, and hence may invalidate the RPO/DT iterator.
2699 std::vector<BasicBlock *> BBVect;
2700 BBVect.reserve(256);
2701 // Needed for value numbering with phi construction to work.
2702 ReversePostOrderTraversal<Function *> RPOT(&F);
2703 for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
2706 BBVect.push_back(*RI);
2708 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2710 Changed |= processBlock(*I);
2715 void GVN::cleanupGlobalSets() {
2717 LeaderTable.clear();
2718 TableAllocator.Reset();
2721 /// verifyRemoved - Verify that the specified instruction does not occur in our
2722 /// internal data structures.
2723 void GVN::verifyRemoved(const Instruction *Inst) const {
2724 VN.verifyRemoved(Inst);
2726 // Walk through the value number scope to make sure the instruction isn't
2727 // ferreted away in it.
2728 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2729 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2730 const LeaderTableEntry *Node = &I->second;
2731 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2733 while (Node->Next) {
2735 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2740 // BB is declared dead, which implied other blocks become dead as well. This
2741 // function is to add all these blocks to "DeadBlocks". For the dead blocks'
2742 // live successors, update their phi nodes by replacing the operands
2743 // corresponding to dead blocks with UndefVal.
2745 void GVN::addDeadBlock(BasicBlock *BB) {
2746 SmallVector<BasicBlock *, 4> NewDead;
2747 SmallSetVector<BasicBlock *, 4> DF;
2749 NewDead.push_back(BB);
2750 while (!NewDead.empty()) {
2751 BasicBlock *D = NewDead.pop_back_val();
2752 if (DeadBlocks.count(D))
2755 // All blocks dominated by D are dead.
2756 SmallVector<BasicBlock *, 8> Dom;
2757 DT->getDescendants(D, Dom);
2758 DeadBlocks.insert(Dom.begin(), Dom.end());
2760 // Figure out the dominance-frontier(D).
2761 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2762 E = Dom.end(); I != E; I++) {
2764 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2765 BasicBlock *S = *SI;
2766 if (DeadBlocks.count(S))
2769 bool AllPredDead = true;
2770 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2771 if (!DeadBlocks.count(*PI)) {
2772 AllPredDead = false;
2777 // S could be proved dead later on. That is why we don't update phi
2778 // operands at this moment.
2781 // While S is not dominated by D, it is dead by now. This could take
2782 // place if S already have a dead predecessor before D is declared
2784 NewDead.push_back(S);
2790 // For the dead blocks' live successors, update their phi nodes by replacing
2791 // the operands corresponding to dead blocks with UndefVal.
2792 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2795 if (DeadBlocks.count(B))
2798 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2799 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
2800 PE = Preds.end(); PI != PE; PI++) {
2801 BasicBlock *P = *PI;
2803 if (!DeadBlocks.count(P))
2806 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2807 if (BasicBlock *S = splitCriticalEdges(P, B))
2808 DeadBlocks.insert(P = S);
2811 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2812 PHINode &Phi = cast<PHINode>(*II);
2813 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2814 UndefValue::get(Phi.getType()));
2820 // If the given branch is recognized as a foldable branch (i.e. conditional
2821 // branch with constant condition), it will perform following analyses and
2823 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2824 // R be the target of the dead out-coming edge.
2825 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2826 // edge. The result of this step will be {X| X is dominated by R}
2827 // 2) Identify those blocks which haves at least one dead prodecessor. The
2828 // result of this step will be dominance-frontier(R).
2829 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2830 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2832 // Return true iff *NEW* dead code are found.
2833 bool GVN::processFoldableCondBr(BranchInst *BI) {
2834 if (!BI || BI->isUnconditional())
2837 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2841 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2842 BI->getSuccessor(1) : BI->getSuccessor(0);
2843 if (DeadBlocks.count(DeadRoot))
2846 if (!DeadRoot->getSinglePredecessor())
2847 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2849 addDeadBlock(DeadRoot);
2853 // performPRE() will trigger assert if it comes across an instruction without
2854 // associated val-num. As it normally has far more live instructions than dead
2855 // instructions, it makes more sense just to "fabricate" a val-number for the
2856 // dead code than checking if instruction involved is dead or not.
2857 void GVN::assignValNumForDeadCode() {
2858 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2859 E = DeadBlocks.end(); I != E; I++) {
2860 BasicBlock *BB = *I;
2861 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2863 Instruction *Inst = &*II;
2864 unsigned ValNum = VN.lookup_or_add(Inst);
2865 addToLeaderTable(ValNum, Inst, BB);